Predicting Postoperative Complications in Glioblastoma Patients Using Machine Learning Models

Introduction: Glioblastoma multiforme (GBM) is the most aggressive primary brain tumor in adults. Despite advanced treatments, postoperative complications remain common and significantly impact patient outcomes. This study aims to predict such complications using machine learning (ML) models. Method: a retrospective analysis was conducted using GBM patient data from open-access sources (TCIA and Kaggle). Preoperative, intraoperative, and postoperative variables were collected. ML models including Logistic Regression, Random Forest, XGBoost, and Long Short-Term Memory (LSTM) were trained and evaluated using metrics such as AUROC, AUPRC, sensitivity, and specificity. Feature importance was assessed using SHAP values. Results: The study included 498 patients (median age: 55 years; 60 % male). Postoperative complications occurred in 30 % of patients, with infections (15 %), hemorrhage (10 %), and neurological deficits (18 %) being most common. LSTM outperformed other models (AUROC: 0.88; AUPRC: 0.64), especially in Grade IV tumors. Key predictors included low preoperative KPS, eloquent tumor location, subtotal resection, and ICU stay >5 days. Conclusions: ML models, especially deep learning (LSTM), effectively predicted postoperative complications in GBM patients. Their integration into clinical workflows may enhance risk stratification, surgical planning, and patient counseling.